| Abstract: | In this paper, we performed papersurvey of deep learning algorithms andmodels in ST-data mining & forecasting.Leading to experiments, we asked ourselves2 questions: 1. There are several STDM-DLmodels, but how well can they learn, andperform forecasting? And 2. If we have acustom dataset, with its data structurevisualised, which model to be learned fromit? We answered these in 2 experiments, thefirst one was that we run the state-of-the artSTDM-DL models and compare theirmetrics. Majorly, the selected models,trained by either METR-LA or PEMS-BAYdataset, predicted the traffic in both spatialand temporal domains. In the second one,we had a fire-call record dataset of the NewTaipei City (NTPC-Fire 2015-17), andimplemented some simple, yet familiarmodels such as Autoencoders and GANs toreconstruct (predict) a rasterised heatmapand LSTM-RNNs, FBProphet and ARIMAin the temporal models to compareperformance in time series forecasting ofdaily, and weekly, incident frequency. In ourfirst experiment, we found out that somestate-of-the-art models, like ST-METANET,STGCN, and Spacetimeformer, had asimilar metrics, with all of them second onlyto the multi-LSTM. And found out that the“Deepforecast Multi-LSTM” is the besttraffic prediction model to date. In oursecond experiment, surprisingly, for oursmall dataset, the FBProphet modeloutperformed even our best LSTM. And ourbest spatial model to reconstruct (predict) araster heatmap was the VariationalAutoencoder (VAE). Given to thesefindings, we have known how analyse thedata via visualization, and implement correctmodels and architectures for each domain inSTDM task. Finally, we will continue todiscover the method to solve environmentalissues, and provide recommendations forfuture subtests, to point out the futureresearch directions for this fast-growingresearch field. |
